Behavioral History Helps LLM Personas Predict People Better Than Descriptions

Summary

This study examines what information helps large language models act as synthetic survey respondents that predict the later choices of specific people. Researchers compare five conditions: no personal information, demographics, personality traits, cognitive scores, and the respondent's earlier survey answers. The analysis uses a two-wave panel of 845 U.S. adults who completed measures of 14 behavioral biases covering risk, time preferences, overconfidence, and reasoning; items used to score the target bias are withheld when behavioral history is supplied. At the population level, all conditions produce averages near the human mean of 7.1 biases, with synthetic averages ranging from 7.1 to 8.1, but this aggregate similarity hides weaker individual differences. Persona descriptions recover 53% to 67% of the human between-person variation, while behavioral history restores variation to approximately the human level. For individual prediction, description-based personas reach only 7% to 12% of the informedness seen in human test-retest responses, compared with 28% when prior answers are added. The behavioral-history condition performs best across all 17 demographic groups, whereas description-based conditions provide little or no information for some groups. Synthetic responses also show stronger education- and income-related differences than the human responses. The study concludes that past behavior is more useful for predicting an individual than a description of that person's attributes.